# The Geometry of Concepts: Sparse Autoencoder Feature Structure

**Authors:** Yuxiao Li, Eric J. Michaud, David D. Baek, Joshua Engels, Xiaoqing Sun, Max Tegmark  
**arXiv:** [2410.19750](https://arxiv.org/abs/2410.19750)  
**Date:** 2024-10-10  
**Summary page:** https://www.summarizepaper.com/en/paper/2410.19750v1/

## Key points

- Study titled "The Geometry of Concepts: Sparse Autoencoder Feature Structure" by Yuxiao Li, Eric J. Michaud, David D. Baek, Joshua Engels, Xiaoqing Sun, and Max Tegmark
- Three levels of structure identified within concept universes generated by sparse autoencoders:
- Atomic level: Presence of "crystals" resembling parallelograms or trapezoids (e.g., man-woman-king-queen)
- Brain level: Significant spatial modularity with distinct lobes for features like mathematics and coding
- Galaxy scale large-scale structure level: Non-isotropic distribution of feature point cloud with power law distribution of eigenvalues showing steepest slope in middle layers
- Use of linear discriminant analysis to enhance quality of parallelograms and function vectors by eliminating global distractor directions like word length
- Quantification of spatial locality of lobes through various metrics revealing clusters of co-occurring features tend to spatially cluster together more than expected if feature geometry were random

## Summary

, , , , 

In their study titled "The Geometry of Concepts: Sparse Autoencoder Feature Structure," Yuxiao Li, Eric J. Michaud, David D. Baek, Joshua Engels, Xiaoqing Sun, and Max Tegmark delve into the intricate structure of concept universes generated by sparse autoencoders. The researchers identify three levels of structure within these universes. At the atomic level, they observe the presence of "crystals" with faces resembling parallelograms or trapezoids. These crystals are exemplified by well-known concepts such as (man-woman-king-queen). By employing linear discriminant analysis to eliminate global distractor directions like word length, the quality of these parallelograms and associated function vectors is significantly enhanced. Moving to the brain level of intermediate-scale structure, the researchers uncover significant spatial modularity. For instance, features related to mathematics and coding form a distinct "lobe," reminiscent of functional lobes observed in neural fMRI images. Through various metrics, they quantify the spatial locality of these lobes and find that clusters of co-occurring features tend to spatially cluster together more than expected if feature geometry were random. Finally, at the galaxy scale large-scale structure level, it is revealed that the feature point cloud exhibits a non-isotropic distribution with a power law distribution of eigenvalues showing steepest slope in middle layers. Additionally, the researchers analyze how clustering entropy varies across different layers. This comprehensive exploration sheds light on the complex and multi-level structural organization present within concept universes generated by sparse autoencoders. The findings contribute valuable insights into understanding the underlying geometry of concepts and their representations in high-dimensional vector spaces.